1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_RPC
39 #include <liblustre.h>
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_net.h>
44 #include <lu_object.h>
45 #include <lnet/types.h>
46 #include "ptlrpc_internal.h"
48 /* The following are visible and mutable through /sys/module/ptlrpc */
49 int test_req_buffer_pressure = 0;
50 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
51 "set non-zero to put pressure on request buffer pools");
52 CFS_MODULE_PARM(at_min, "i", int, 0644,
53 "Adaptive timeout minimum (sec)");
54 CFS_MODULE_PARM(at_max, "i", int, 0644,
55 "Adaptive timeout maximum (sec)");
56 CFS_MODULE_PARM(at_history, "i", int, 0644,
57 "Adaptive timeouts remember the slowest event that took place "
58 "within this period (sec)");
59 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
60 "How soon before an RPC deadline to send an early reply");
61 CFS_MODULE_PARM(at_extra, "i", int, 0644,
62 "How much extra time to give with each early reply");
66 static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc);
68 static CFS_LIST_HEAD(ptlrpc_all_services);
69 cfs_spinlock_t ptlrpc_all_services_lock;
72 ptlrpc_alloc_request_buffer (int size)
76 if (size > SVC_BUF_VMALLOC_THRESHOLD)
77 OBD_VMALLOC(ptr, size);
85 ptlrpc_free_request_buffer (char *ptr, int size)
87 if (size > SVC_BUF_VMALLOC_THRESHOLD)
93 struct ptlrpc_request_buffer_desc *
94 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
96 struct ptlrpc_request_buffer_desc *rqbd;
102 rqbd->rqbd_service = svc;
103 rqbd->rqbd_refcount = 0;
104 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
105 rqbd->rqbd_cbid.cbid_arg = rqbd;
106 CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
107 rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size);
109 if (rqbd->rqbd_buffer == NULL) {
114 cfs_spin_lock(&svc->srv_lock);
115 cfs_list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
117 cfs_spin_unlock(&svc->srv_lock);
123 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
125 struct ptlrpc_service *svc = rqbd->rqbd_service;
127 LASSERT (rqbd->rqbd_refcount == 0);
128 LASSERT (cfs_list_empty(&rqbd->rqbd_reqs));
130 cfs_spin_lock(&svc->srv_lock);
131 cfs_list_del(&rqbd->rqbd_list);
133 cfs_spin_unlock(&svc->srv_lock);
135 ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size);
140 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
142 struct ptlrpc_request_buffer_desc *rqbd;
145 CDEBUG(D_RPCTRACE, "%s: allocate %d new %d-byte reqbufs (%d/%d left)\n",
146 svc->srv_name, svc->srv_nbuf_per_group, svc->srv_buf_size,
147 svc->srv_nrqbd_receiving, svc->srv_nbufs);
148 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
149 rqbd = ptlrpc_alloc_rqbd(svc);
152 CERROR ("%s: Can't allocate request buffer\n",
157 if (ptlrpc_server_post_idle_rqbds(svc) < 0)
165 * Part of Rep-Ack logic.
166 * Puts a lock and its mode into reply state assotiated to request reply.
169 ptlrpc_save_lock(struct ptlrpc_request *req,
170 struct lustre_handle *lock, int mode, int no_ack)
172 struct ptlrpc_reply_state *rs = req->rq_reply_state;
176 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
178 if (req->rq_export->exp_disconnected) {
179 ldlm_lock_decref(lock, mode);
181 idx = rs->rs_nlocks++;
182 rs->rs_locks[idx] = *lock;
183 rs->rs_modes[idx] = mode;
184 rs->rs_difficult = 1;
185 rs->rs_no_ack = !!no_ack;
191 #define HRT_RUNNING 0
192 #define HRT_STOPPING 1
194 struct ptlrpc_hr_thread {
195 cfs_spinlock_t hrt_lock;
196 unsigned long hrt_flags;
197 cfs_waitq_t hrt_wait;
198 cfs_list_t hrt_queue;
199 cfs_completion_t hrt_completion;
202 struct ptlrpc_hr_service {
206 struct ptlrpc_hr_thread hr_threads[0];
210 cfs_list_t rsb_replies;
211 struct ptlrpc_service *rsb_svc;
212 unsigned int rsb_n_replies;
216 * A pointer to per-node reply handling service.
218 static struct ptlrpc_hr_service *ptlrpc_hr = NULL;
221 * maximum mumber of replies scheduled in one batch
223 #define MAX_SCHEDULED 256
226 * Initialize a reply batch.
230 static void rs_batch_init(struct rs_batch *b)
232 memset(b, 0, sizeof *b);
233 CFS_INIT_LIST_HEAD(&b->rsb_replies);
237 * Choose an hr thread to dispatch requests to.
239 static unsigned int get_hr_thread_index(struct ptlrpc_hr_service *hr)
243 /* Concurrent modification of hr_index w/o any spinlock
244 protection is harmless as long as the result fits
245 [0..(hr_n_threads-1)] range and each thread gets near equal
248 hr->hr_index = (idx >= hr->hr_n_threads - 1) ? 0 : idx + 1;
253 * Dispatch all replies accumulated in the batch to one from
254 * dedicated reply handling threads.
258 static void rs_batch_dispatch(struct rs_batch *b)
260 if (b->rsb_n_replies != 0) {
261 struct ptlrpc_hr_service *hr = ptlrpc_hr;
264 idx = get_hr_thread_index(hr);
266 cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
267 cfs_list_splice_init(&b->rsb_replies,
268 &hr->hr_threads[idx].hrt_queue);
269 cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
270 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
271 b->rsb_n_replies = 0;
276 * Add a reply to a batch.
277 * Add one reply object to a batch, schedule batched replies if overload.
282 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
284 struct ptlrpc_service *svc = rs->rs_service;
286 if (svc != b->rsb_svc || b->rsb_n_replies >= MAX_SCHEDULED) {
287 if (b->rsb_svc != NULL) {
288 rs_batch_dispatch(b);
289 cfs_spin_unlock(&b->rsb_svc->srv_lock);
291 cfs_spin_lock(&svc->srv_lock);
294 cfs_spin_lock(&rs->rs_lock);
295 rs->rs_scheduled_ever = 1;
296 if (rs->rs_scheduled == 0) {
297 cfs_list_move(&rs->rs_list, &b->rsb_replies);
298 rs->rs_scheduled = 1;
301 rs->rs_committed = 1;
302 cfs_spin_unlock(&rs->rs_lock);
306 * Reply batch finalization.
307 * Dispatch remaining replies from the batch
308 * and release remaining spinlock.
312 static void rs_batch_fini(struct rs_batch *b)
314 if (b->rsb_svc != 0) {
315 rs_batch_dispatch(b);
316 cfs_spin_unlock(&b->rsb_svc->srv_lock);
320 #define DECLARE_RS_BATCH(b) struct rs_batch b
322 #else /* __KERNEL__ */
324 #define rs_batch_init(b) do{}while(0)
325 #define rs_batch_fini(b) do{}while(0)
326 #define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
327 #define DECLARE_RS_BATCH(b)
329 #endif /* __KERNEL__ */
332 * Put reply state into a queue for processing because we received
333 * ACK from the client
335 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
338 struct ptlrpc_hr_service *hr = ptlrpc_hr;
342 LASSERT(cfs_list_empty(&rs->rs_list));
344 idx = get_hr_thread_index(hr);
345 cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
346 cfs_list_add_tail(&rs->rs_list, &hr->hr_threads[idx].hrt_queue);
347 cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
348 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
351 cfs_list_add_tail(&rs->rs_list, &rs->rs_service->srv_reply_queue);
356 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
360 LASSERT_SPIN_LOCKED(&rs->rs_service->srv_lock);
361 LASSERT_SPIN_LOCKED(&rs->rs_lock);
362 LASSERT (rs->rs_difficult);
363 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
365 if (rs->rs_scheduled) { /* being set up or already notified */
370 rs->rs_scheduled = 1;
371 cfs_list_del_init(&rs->rs_list);
372 ptlrpc_dispatch_difficult_reply(rs);
376 void ptlrpc_commit_replies(struct obd_export *exp)
378 struct ptlrpc_reply_state *rs, *nxt;
379 DECLARE_RS_BATCH(batch);
382 rs_batch_init(&batch);
383 /* Find any replies that have been committed and get their service
384 * to attend to complete them. */
386 /* CAVEAT EMPTOR: spinlock ordering!!! */
387 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
388 cfs_list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
390 LASSERT (rs->rs_difficult);
391 /* VBR: per-export last_committed */
392 LASSERT(rs->rs_export);
393 if (rs->rs_transno <= exp->exp_last_committed) {
394 cfs_list_del_init(&rs->rs_obd_list);
395 rs_batch_add(&batch, rs);
398 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
399 rs_batch_fini(&batch);
404 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
406 struct ptlrpc_request_buffer_desc *rqbd;
411 cfs_spin_lock(&svc->srv_lock);
413 if (cfs_list_empty (&svc->srv_idle_rqbds)) {
414 cfs_spin_unlock(&svc->srv_lock);
418 rqbd = cfs_list_entry(svc->srv_idle_rqbds.next,
419 struct ptlrpc_request_buffer_desc,
421 cfs_list_del (&rqbd->rqbd_list);
423 /* assume we will post successfully */
424 svc->srv_nrqbd_receiving++;
425 cfs_list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
427 cfs_spin_unlock(&svc->srv_lock);
429 rc = ptlrpc_register_rqbd(rqbd);
436 cfs_spin_lock(&svc->srv_lock);
438 svc->srv_nrqbd_receiving--;
439 cfs_list_del(&rqbd->rqbd_list);
440 cfs_list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
442 /* Don't complain if no request buffers are posted right now; LNET
443 * won't drop requests because we set the portal lazy! */
445 cfs_spin_unlock(&svc->srv_lock);
451 * Start a service with parameters from struct ptlrpc_service_conf \a c
452 * as opposed to directly calling ptlrpc_init_svc with tons of arguments.
454 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
455 svc_handler_t h, char *name,
456 struct proc_dir_entry *proc_entry,
457 svcreq_printfn_t prntfn,
460 return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
461 c->psc_max_req_size, c->psc_max_reply_size,
462 c->psc_req_portal, c->psc_rep_portal,
463 c->psc_watchdog_factor,
465 prntfn, c->psc_min_threads, c->psc_max_threads,
466 threadname, c->psc_ctx_tags, NULL);
468 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
470 static void ptlrpc_at_timer(unsigned long castmeharder)
472 struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
473 svc->srv_at_check = 1;
474 svc->srv_at_checktime = cfs_time_current();
475 cfs_waitq_signal(&svc->srv_waitq);
479 * Initialize service on a given portal.
480 * This includes starting serving threads , allocating and posting rqbds and
482 * \a nbufs is how many buffers to post
483 * \a bufsize is buffer size to post
484 * \a max_req_size - maximum request size to be accepted for this service
485 * \a max_reply_size maximum reply size this service can ever send
486 * \a req_portal - portal to listed for requests on
487 * \a rep_portal - portal of where to send replies to
488 * \a watchdog_factor soft watchdog timeout multiplifier to print stuck service traces.
489 * \a handler - function to process every new request
490 * \a name - service name
491 * \a proc_entry - entry in the /proc tree for sttistics reporting
492 * \a min_threads \a max_threads - min/max number of service threads to start.
493 * \a threadname should be 11 characters or less - 3 will be added on
494 * \a hp_handler - function to determine priority of the request, also called
495 * on every new request.
497 struct ptlrpc_service *
498 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
499 int req_portal, int rep_portal, int watchdog_factor,
500 svc_handler_t handler, char *name,
501 cfs_proc_dir_entry_t *proc_entry,
502 svcreq_printfn_t svcreq_printfn,
503 int min_threads, int max_threads,
504 char *threadname, __u32 ctx_tags,
505 svc_hpreq_handler_t hp_handler)
508 struct ptlrpc_at_array *array;
509 struct ptlrpc_service *service;
510 unsigned int size, index;
514 LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
515 LASSERT (ctx_tags != 0);
517 OBD_ALLOC_PTR(service);
521 /* First initialise enough for early teardown */
523 service->srv_name = name;
524 cfs_spin_lock_init(&service->srv_lock);
525 CFS_INIT_LIST_HEAD(&service->srv_threads);
526 cfs_waitq_init(&service->srv_waitq);
528 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
529 service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
530 service->srv_buf_size = bufsize;
531 service->srv_rep_portal = rep_portal;
532 service->srv_req_portal = req_portal;
533 service->srv_watchdog_factor = watchdog_factor;
534 service->srv_handler = handler;
535 service->srv_request_history_print_fn = svcreq_printfn;
536 service->srv_request_seq = 1; /* valid seq #s start at 1 */
537 service->srv_request_max_cull_seq = 0;
538 service->srv_threads_min = min_threads;
539 service->srv_threads_max = max_threads;
540 service->srv_thread_name = threadname;
541 service->srv_ctx_tags = ctx_tags;
542 service->srv_hpreq_handler = hp_handler;
543 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
544 service->srv_hpreq_count = 0;
545 service->srv_n_hpreq = 0;
547 rc = LNetSetLazyPortal(service->srv_req_portal);
550 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
551 CFS_INIT_LIST_HEAD(&service->srv_request_hpq);
552 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
553 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
554 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
555 CFS_INIT_LIST_HEAD(&service->srv_request_history);
556 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
558 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
560 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
561 cfs_waitq_init(&service->srv_free_rs_waitq);
562 cfs_atomic_set(&service->srv_n_difficult_replies, 0);
564 cfs_spin_lock_init(&service->srv_at_lock);
565 CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
567 array = &service->srv_at_array;
568 size = at_est2timeout(at_max);
569 array->paa_size = size;
570 array->paa_count = 0;
571 array->paa_deadline = -1;
573 /* allocate memory for srv_at_array (ptlrpc_at_array) */
574 OBD_ALLOC(array->paa_reqs_array, sizeof(cfs_list_t) * size);
575 if (array->paa_reqs_array == NULL)
578 for (index = 0; index < size; index++)
579 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
581 OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size);
582 if (array->paa_reqs_count == NULL)
585 cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
586 /* At SOW, service time should be quick; 10s seems generous. If client
587 timeout is less than this, we'll be sending an early reply. */
588 at_init(&service->srv_at_estimate, 10, 0);
590 cfs_spin_lock (&ptlrpc_all_services_lock);
591 cfs_list_add (&service->srv_list, &ptlrpc_all_services);
592 cfs_spin_unlock (&ptlrpc_all_services_lock);
594 /* Now allocate the request buffers */
595 rc = ptlrpc_grow_req_bufs(service);
596 /* We shouldn't be under memory pressure at startup, so
597 * fail if we can't post all our buffers at this time. */
601 /* Now allocate pool of reply buffers */
602 /* Increase max reply size to next power of two */
603 service->srv_max_reply_size = 1;
604 while (service->srv_max_reply_size <
605 max_reply_size + SPTLRPC_MAX_PAYLOAD)
606 service->srv_max_reply_size <<= 1;
608 if (proc_entry != NULL)
609 ptlrpc_lprocfs_register_service(proc_entry, service);
611 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
612 service->srv_name, service->srv_req_portal);
616 ptlrpc_unregister_service(service);
621 * to actually free the request, must be called without holding svc_lock.
622 * note it's caller's responsibility to unlink req->rq_list.
624 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
626 LASSERT(cfs_atomic_read(&req->rq_refcount) == 0);
627 LASSERT(cfs_list_empty(&req->rq_timed_list));
629 /* DEBUG_REQ() assumes the reply state of a request with a valid
630 * ref will not be destroyed until that reference is dropped. */
631 ptlrpc_req_drop_rs(req);
633 sptlrpc_svc_ctx_decref(req);
635 if (req != &req->rq_rqbd->rqbd_req) {
636 /* NB request buffers use an embedded
637 * req if the incoming req unlinked the
638 * MD; this isn't one of them! */
639 OBD_FREE(req, sizeof(*req));
644 * increment the number of active requests consuming service threads.
646 void ptlrpc_server_active_request_inc(struct ptlrpc_request *req)
648 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
649 struct ptlrpc_service *svc = rqbd->rqbd_service;
651 cfs_spin_lock(&svc->srv_lock);
652 svc->srv_n_active_reqs++;
653 cfs_spin_unlock(&svc->srv_lock);
657 * decrement the number of active requests consuming service threads.
659 void ptlrpc_server_active_request_dec(struct ptlrpc_request *req)
661 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
662 struct ptlrpc_service *svc = rqbd->rqbd_service;
664 cfs_spin_lock(&svc->srv_lock);
665 svc->srv_n_active_reqs--;
666 cfs_spin_unlock(&svc->srv_lock);
670 * drop a reference count of the request. if it reaches 0, we either
671 * put it into history list, or free it immediately.
673 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
675 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
676 struct ptlrpc_service *svc = rqbd->rqbd_service;
681 if (!cfs_atomic_dec_and_test(&req->rq_refcount))
684 cfs_spin_lock(&svc->srv_at_lock);
685 if (req->rq_at_linked) {
686 struct ptlrpc_at_array *array = &svc->srv_at_array;
687 __u32 index = req->rq_at_index;
689 LASSERT(!cfs_list_empty(&req->rq_timed_list));
690 cfs_list_del_init(&req->rq_timed_list);
691 cfs_spin_lock(&req->rq_lock);
692 req->rq_at_linked = 0;
693 cfs_spin_unlock(&req->rq_lock);
694 array->paa_reqs_count[index]--;
697 LASSERT(cfs_list_empty(&req->rq_timed_list));
698 cfs_spin_unlock(&svc->srv_at_lock);
700 /* finalize request */
701 if (req->rq_export) {
702 class_export_put(req->rq_export);
703 req->rq_export = NULL;
706 cfs_spin_lock(&svc->srv_lock);
708 svc->srv_n_active_reqs--;
709 cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs);
711 refcount = --(rqbd->rqbd_refcount);
713 /* request buffer is now idle: add to history */
714 cfs_list_del(&rqbd->rqbd_list);
715 cfs_list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
716 svc->srv_n_history_rqbds++;
718 /* cull some history?
719 * I expect only about 1 or 2 rqbds need to be recycled here */
720 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
721 rqbd = cfs_list_entry(svc->srv_history_rqbds.next,
722 struct ptlrpc_request_buffer_desc,
725 cfs_list_del(&rqbd->rqbd_list);
726 svc->srv_n_history_rqbds--;
728 /* remove rqbd's reqs from svc's req history while
729 * I've got the service lock */
730 cfs_list_for_each(tmp, &rqbd->rqbd_reqs) {
731 req = cfs_list_entry(tmp, struct ptlrpc_request,
733 /* Track the highest culled req seq */
734 if (req->rq_history_seq >
735 svc->srv_request_max_cull_seq)
736 svc->srv_request_max_cull_seq =
738 cfs_list_del(&req->rq_history_list);
741 cfs_spin_unlock(&svc->srv_lock);
743 cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
744 req = cfs_list_entry(rqbd->rqbd_reqs.next,
745 struct ptlrpc_request,
747 cfs_list_del(&req->rq_list);
748 ptlrpc_server_free_request(req);
751 cfs_spin_lock(&svc->srv_lock);
753 * now all reqs including the embedded req has been
754 * disposed, schedule request buffer for re-use.
756 LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) ==
758 cfs_list_add_tail(&rqbd->rqbd_list,
759 &svc->srv_idle_rqbds);
762 cfs_spin_unlock(&svc->srv_lock);
763 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
764 /* If we are low on memory, we are not interested in history */
765 cfs_list_del(&req->rq_list);
766 cfs_list_del_init(&req->rq_history_list);
767 cfs_spin_unlock(&svc->srv_lock);
769 ptlrpc_server_free_request(req);
771 cfs_spin_unlock(&svc->srv_lock);
776 * to finish a request: stop sending more early replies, and release
777 * the request. should be called after we finished handling the request.
779 static void ptlrpc_server_finish_request(struct ptlrpc_request *req)
781 ptlrpc_server_drop_request(req);
785 * This function makes sure dead exports are evicted in a timely manner.
786 * This function is only called when some export receives a message (i.e.,
787 * the network is up.)
789 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
791 struct obd_export *oldest_exp;
792 time_t oldest_time, new_time;
798 /* Compensate for slow machines, etc, by faking our request time
799 into the future. Although this can break the strict time-ordering
800 of the list, we can be really lazy here - we don't have to evict
801 at the exact right moment. Eventually, all silent exports
802 will make it to the top of the list. */
804 /* Do not pay attention on 1sec or smaller renewals. */
805 new_time = cfs_time_current_sec() + extra_delay;
806 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
809 exp->exp_last_request_time = new_time;
810 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
811 exp->exp_client_uuid.uuid,
812 exp->exp_last_request_time, exp);
814 /* exports may get disconnected from the chain even though the
815 export has references, so we must keep the spin lock while
816 manipulating the lists */
817 cfs_spin_lock(&exp->exp_obd->obd_dev_lock);
819 if (cfs_list_empty(&exp->exp_obd_chain_timed)) {
820 /* this one is not timed */
821 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
825 cfs_list_move_tail(&exp->exp_obd_chain_timed,
826 &exp->exp_obd->obd_exports_timed);
828 oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next,
829 struct obd_export, exp_obd_chain_timed);
830 oldest_time = oldest_exp->exp_last_request_time;
831 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
833 if (exp->exp_obd->obd_recovering) {
834 /* be nice to everyone during recovery */
839 /* Note - racing to start/reset the obd_eviction timer is safe */
840 if (exp->exp_obd->obd_eviction_timer == 0) {
841 /* Check if the oldest entry is expired. */
842 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
844 /* We need a second timer, in case the net was down and
845 * it just came back. Since the pinger may skip every
846 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
847 * we better wait for 3. */
848 exp->exp_obd->obd_eviction_timer =
849 cfs_time_current_sec() + 3 * PING_INTERVAL;
850 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
851 exp->exp_obd->obd_name,
852 obd_export_nid2str(oldest_exp), oldest_time);
855 if (cfs_time_current_sec() >
856 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
857 /* The evictor won't evict anyone who we've heard from
858 * recently, so we don't have to check before we start
860 if (!ping_evictor_wake(exp))
861 exp->exp_obd->obd_eviction_timer = 0;
869 * Sanity check request \a req.
870 * Return 0 if all is ok, error code otherwise.
872 static int ptlrpc_check_req(struct ptlrpc_request *req)
874 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
875 req->rq_export->exp_conn_cnt)) {
876 DEBUG_REQ(D_ERROR, req,
877 "DROPPING req from old connection %d < %d",
878 lustre_msg_get_conn_cnt(req->rq_reqmsg),
879 req->rq_export->exp_conn_cnt);
882 if (unlikely(req->rq_export->exp_obd &&
883 req->rq_export->exp_obd->obd_fail)) {
884 /* Failing over, don't handle any more reqs, send
885 error response instead. */
886 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
887 req, req->rq_export->exp_obd->obd_name);
888 req->rq_status = -ENODEV;
896 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
898 struct ptlrpc_at_array *array = &svc->srv_at_array;
901 cfs_spin_lock(&svc->srv_at_lock);
902 if (array->paa_count == 0) {
903 cfs_timer_disarm(&svc->srv_at_timer);
904 cfs_spin_unlock(&svc->srv_at_lock);
908 /* Set timer for closest deadline */
909 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
912 ptlrpc_at_timer((unsigned long)svc);
914 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
915 cfs_spin_unlock(&svc->srv_at_lock);
916 CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
919 /* Add rpc to early reply check list */
920 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
922 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
923 struct ptlrpc_request *rq = NULL;
924 struct ptlrpc_at_array *array = &svc->srv_at_array;
931 if (req->rq_no_reply)
934 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
937 cfs_spin_lock(&svc->srv_at_lock);
938 LASSERT(cfs_list_empty(&req->rq_timed_list));
940 index = (unsigned long)req->rq_deadline % array->paa_size;
941 if (array->paa_reqs_count[index] > 0) {
942 /* latest rpcs will have the latest deadlines in the list,
943 * so search backward. */
944 cfs_list_for_each_entry_reverse(rq,
945 &array->paa_reqs_array[index],
947 if (req->rq_deadline >= rq->rq_deadline) {
948 cfs_list_add(&req->rq_timed_list,
955 /* Add the request at the head of the list */
956 if (cfs_list_empty(&req->rq_timed_list))
957 cfs_list_add(&req->rq_timed_list,
958 &array->paa_reqs_array[index]);
960 cfs_spin_lock(&req->rq_lock);
961 req->rq_at_linked = 1;
962 cfs_spin_unlock(&req->rq_lock);
963 req->rq_at_index = index;
964 array->paa_reqs_count[index]++;
966 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
967 array->paa_deadline = req->rq_deadline;
970 cfs_spin_unlock(&svc->srv_at_lock);
973 ptlrpc_at_set_timer(svc);
978 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
980 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
981 struct ptlrpc_request *reqcopy;
982 struct lustre_msg *reqmsg;
983 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
988 /* deadline is when the client expects us to reply, margin is the
989 difference between clients' and servers' expectations */
990 DEBUG_REQ(D_ADAPTTO, req,
991 "%ssending early reply (deadline %+lds, margin %+lds) for "
992 "%d+%d", AT_OFF ? "AT off - not " : "",
993 olddl, olddl - at_get(&svc->srv_at_estimate),
994 at_get(&svc->srv_at_estimate), at_extra);
1000 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1001 "not sending early reply. Consider increasing "
1002 "at_early_margin (%d)?", olddl, at_early_margin);
1004 /* Return an error so we're not re-added to the timed list. */
1008 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1009 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1010 "but no AT support");
1014 if (req->rq_export &&
1015 lustre_msg_get_flags(req->rq_reqmsg) &
1016 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1017 /* During recovery, we don't want to send too many early
1018 * replies, but on the other hand we want to make sure the
1019 * client has enough time to resend if the rpc is lost. So
1020 * during the recovery period send at least 4 early replies,
1021 * spacing them every at_extra if we can. at_estimate should
1022 * always equal this fixed value during recovery. */
1023 at_measured(&svc->srv_at_estimate, min(at_extra,
1024 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1026 /* Fake our processing time into the future to ask the clients
1027 * for some extra amount of time */
1028 at_measured(&svc->srv_at_estimate, at_extra +
1029 cfs_time_current_sec() -
1030 req->rq_arrival_time.tv_sec);
1032 /* Check to see if we've actually increased the deadline -
1033 * we may be past adaptive_max */
1034 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1035 at_get(&svc->srv_at_estimate)) {
1036 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1037 "(%ld/%ld), not sending early reply\n",
1038 olddl, req->rq_arrival_time.tv_sec +
1039 at_get(&svc->srv_at_estimate) -
1040 cfs_time_current_sec());
1044 newdl = cfs_time_current_sec() + at_get(&svc->srv_at_estimate);
1046 OBD_ALLOC(reqcopy, sizeof *reqcopy);
1047 if (reqcopy == NULL)
1049 OBD_ALLOC(reqmsg, req->rq_reqlen);
1051 OBD_FREE(reqcopy, sizeof *reqcopy);
1056 reqcopy->rq_reply_state = NULL;
1057 reqcopy->rq_rep_swab_mask = 0;
1058 reqcopy->rq_pack_bulk = 0;
1059 reqcopy->rq_pack_udesc = 0;
1060 reqcopy->rq_packed_final = 0;
1061 sptlrpc_svc_ctx_addref(reqcopy);
1062 /* We only need the reqmsg for the magic */
1063 reqcopy->rq_reqmsg = reqmsg;
1064 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1066 LASSERT(cfs_atomic_read(&req->rq_refcount));
1067 /** if it is last refcount then early reply isn't needed */
1068 if (cfs_atomic_read(&req->rq_refcount) == 1) {
1069 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1070 "abort sending early reply\n");
1071 GOTO(out, rc = -EINVAL);
1074 /* Connection ref */
1075 reqcopy->rq_export = class_conn2export(
1076 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1077 if (reqcopy->rq_export == NULL)
1078 GOTO(out, rc = -ENODEV);
1081 class_export_rpc_get(reqcopy->rq_export);
1082 if (reqcopy->rq_export->exp_obd &&
1083 reqcopy->rq_export->exp_obd->obd_fail)
1084 GOTO(out_put, rc = -ENODEV);
1086 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1090 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1093 /* Adjust our own deadline to what we told the client */
1094 req->rq_deadline = newdl;
1095 req->rq_early_count++; /* number sent, server side */
1097 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1100 /* Free the (early) reply state from lustre_pack_reply.
1101 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1102 ptlrpc_req_drop_rs(reqcopy);
1105 class_export_rpc_put(reqcopy->rq_export);
1106 class_export_put(reqcopy->rq_export);
1108 sptlrpc_svc_ctx_decref(reqcopy);
1109 OBD_FREE(reqmsg, req->rq_reqlen);
1110 OBD_FREE(reqcopy, sizeof *reqcopy);
1114 /* Send early replies to everybody expiring within at_early_margin
1115 asking for at_extra time */
1116 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
1118 struct ptlrpc_request *rq, *n;
1119 cfs_list_t work_list;
1120 struct ptlrpc_at_array *array = &svc->srv_at_array;
1123 time_t now = cfs_time_current_sec();
1124 cfs_duration_t delay;
1125 int first, counter = 0;
1128 cfs_spin_lock(&svc->srv_at_lock);
1129 if (svc->srv_at_check == 0) {
1130 cfs_spin_unlock(&svc->srv_at_lock);
1133 delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
1134 svc->srv_at_check = 0;
1136 if (array->paa_count == 0) {
1137 cfs_spin_unlock(&svc->srv_at_lock);
1141 /* The timer went off, but maybe the nearest rpc already completed. */
1142 first = array->paa_deadline - now;
1143 if (first > at_early_margin) {
1144 /* We've still got plenty of time. Reset the timer. */
1145 cfs_spin_unlock(&svc->srv_at_lock);
1146 ptlrpc_at_set_timer(svc);
1150 /* We're close to a timeout, and we don't know how much longer the
1151 server will take. Send early replies to everyone expiring soon. */
1152 CFS_INIT_LIST_HEAD(&work_list);
1154 index = (unsigned long)array->paa_deadline % array->paa_size;
1155 count = array->paa_count;
1157 count -= array->paa_reqs_count[index];
1158 cfs_list_for_each_entry_safe(rq, n,
1159 &array->paa_reqs_array[index],
1161 if (rq->rq_deadline <= now + at_early_margin) {
1162 cfs_list_del_init(&rq->rq_timed_list);
1164 * ptlrpc_server_drop_request() may drop
1165 * refcount to 0 already. Let's check this and
1166 * don't add entry to work_list
1168 if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1169 cfs_list_add(&rq->rq_timed_list, &work_list);
1171 array->paa_reqs_count[index]--;
1173 cfs_spin_lock(&rq->rq_lock);
1174 rq->rq_at_linked = 0;
1175 cfs_spin_unlock(&rq->rq_lock);
1179 /* update the earliest deadline */
1180 if (deadline == -1 || rq->rq_deadline < deadline)
1181 deadline = rq->rq_deadline;
1186 if (++index >= array->paa_size)
1189 array->paa_deadline = deadline;
1190 cfs_spin_unlock(&svc->srv_at_lock);
1192 /* we have a new earliest deadline, restart the timer */
1193 ptlrpc_at_set_timer(svc);
1195 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1196 "replies\n", first, at_extra, counter);
1198 /* We're already past request deadlines before we even get a
1199 chance to send early replies */
1200 LCONSOLE_WARN("%s: This server is not able to keep up with "
1201 "request traffic (cpu-bound).\n", svc->srv_name);
1202 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1203 "delay="CFS_DURATION_T"(jiff)\n",
1204 counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
1205 at_get(&svc->srv_at_estimate), delay);
1208 /* we took additional refcount so entries can't be deleted from list, no
1209 * locking is needed */
1210 while (!cfs_list_empty(&work_list)) {
1211 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1213 cfs_list_del_init(&rq->rq_timed_list);
1215 if (ptlrpc_at_send_early_reply(rq) == 0)
1216 ptlrpc_at_add_timed(rq);
1218 ptlrpc_server_drop_request(rq);
1225 * Put the request to the export list if the request may become
1226 * a high priority one.
1228 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
1229 struct ptlrpc_request *req)
1234 if (svc->srv_hpreq_handler) {
1235 rc = svc->srv_hpreq_handler(req);
1239 if (req->rq_export && req->rq_ops) {
1240 cfs_spin_lock(&req->rq_export->exp_lock);
1241 cfs_list_add(&req->rq_exp_list,
1242 &req->rq_export->exp_queued_rpc);
1243 cfs_spin_unlock(&req->rq_export->exp_lock);
1249 /** Remove the request from the export list. */
1250 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1253 if (req->rq_export && req->rq_ops) {
1254 cfs_spin_lock(&req->rq_export->exp_lock);
1255 cfs_list_del_init(&req->rq_exp_list);
1256 cfs_spin_unlock(&req->rq_export->exp_lock);
1262 * Make the request a high priority one.
1264 * All the high priority requests are queued in a separate FIFO
1265 * ptlrpc_service::srv_request_hpq list which is parallel to
1266 * ptlrpc_service::srv_request_queue list but has a higher priority
1269 * \see ptlrpc_server_handle_request().
1271 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
1272 struct ptlrpc_request *req)
1275 LASSERT(svc != NULL);
1276 cfs_spin_lock(&req->rq_lock);
1277 if (req->rq_hp == 0) {
1278 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1280 /* Add to the high priority queue. */
1281 cfs_list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1283 if (opc != OBD_PING)
1284 DEBUG_REQ(D_NET, req, "high priority req");
1286 cfs_spin_unlock(&req->rq_lock);
1291 * \see ptlrpc_hpreq_reorder_nolock
1293 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1295 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1298 cfs_spin_lock(&svc->srv_lock);
1299 /* It may happen that the request is already taken for the processing
1300 * but still in the export list, do not re-add it into the HP list. */
1301 if (req->rq_phase == RQ_PHASE_NEW)
1302 ptlrpc_hpreq_reorder_nolock(svc, req);
1303 cfs_spin_unlock(&svc->srv_lock);
1307 /** Check if the request is a high priority one. */
1308 static int ptlrpc_server_hpreq_check(struct ptlrpc_request *req)
1313 /* Check by request opc. */
1314 opc = lustre_msg_get_opc(req->rq_reqmsg);
1315 if (opc == OBD_PING)
1318 /* Perform request specific check. */
1319 if (req->rq_ops && req->rq_ops->hpreq_check)
1320 rc = req->rq_ops->hpreq_check(req);
1324 /** Check if a request is a high priority one. */
1325 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1326 struct ptlrpc_request *req)
1331 rc = ptlrpc_server_hpreq_check(req);
1335 cfs_spin_lock(&svc->srv_lock);
1336 /* Before inserting the request into the queue, check if it is not
1337 * inserted yet, or even already handled -- it may happen due to
1338 * a racing ldlm_server_blocking_ast(). */
1339 if (req->rq_phase == RQ_PHASE_NEW && cfs_list_empty(&req->rq_list)) {
1341 ptlrpc_hpreq_reorder_nolock(svc, req);
1343 cfs_list_add_tail(&req->rq_list,
1344 &svc->srv_request_queue);
1346 cfs_spin_unlock(&svc->srv_lock);
1352 * Only allow normal priority requests on a service that has a high-priority
1353 * queue if forced (i.e. cleanup), if there are other high priority requests
1354 * already being processed (i.e. those threads can service more high-priority
1355 * requests), or if there are enough idle threads that a later thread can do
1356 * a high priority request.
1358 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1360 return force || !svc->srv_hpreq_handler || svc->srv_n_hpreq > 0 ||
1361 svc->srv_threads_running <= svc->srv_threads_started - 2;
1365 * Fetch a request for processing from queue of unprocessed requests.
1366 * Favors high-priority requests.
1367 * Returns a pointer to fetched request.
1369 static struct ptlrpc_request *
1370 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1372 struct ptlrpc_request *req = NULL;
1375 if (ptlrpc_server_allow_normal(svc, force) &&
1376 !cfs_list_empty(&svc->srv_request_queue) &&
1377 (cfs_list_empty(&svc->srv_request_hpq) ||
1378 svc->srv_hpreq_count >= svc->srv_hpreq_ratio)) {
1379 req = cfs_list_entry(svc->srv_request_queue.next,
1380 struct ptlrpc_request, rq_list);
1381 svc->srv_hpreq_count = 0;
1382 } else if (!cfs_list_empty(&svc->srv_request_hpq)) {
1383 req = cfs_list_entry(svc->srv_request_hpq.next,
1384 struct ptlrpc_request, rq_list);
1385 svc->srv_hpreq_count++;
1391 * Returns true if there are requests available in incoming
1392 * request queue for processing and it is allowed to fetch them
1393 * \see ptlrpc_server_allow_normal
1395 static int ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1397 return ((ptlrpc_server_allow_normal(svc, force) &&
1398 !cfs_list_empty(&svc->srv_request_queue)) ||
1399 !cfs_list_empty(&svc->srv_request_hpq));
1403 * Handle freshly incoming reqs, add to timed early reply list,
1404 * pass on to regular request queue.
1405 * All incoming requests pass through here before getting into
1406 * ptlrpc_server_handle_req later on.
1409 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1411 struct ptlrpc_request *req;
1418 cfs_spin_lock(&svc->srv_lock);
1419 if (cfs_list_empty(&svc->srv_req_in_queue)) {
1420 cfs_spin_unlock(&svc->srv_lock);
1424 req = cfs_list_entry(svc->srv_req_in_queue.next,
1425 struct ptlrpc_request, rq_list);
1426 cfs_list_del_init (&req->rq_list);
1427 /* Consider this still a "queued" request as far as stats are
1429 cfs_spin_unlock(&svc->srv_lock);
1431 /* go through security check/transform */
1432 rc = sptlrpc_svc_unwrap_request(req);
1436 case SECSVC_COMPLETE:
1437 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1446 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1447 * redo it wouldn't be harmful.
1449 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1450 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1452 CERROR("error unpacking request: ptl %d from %s "
1453 "x"LPU64"\n", svc->srv_req_portal,
1454 libcfs_id2str(req->rq_peer), req->rq_xid);
1459 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1461 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1462 LPU64"\n", svc->srv_req_portal,
1463 libcfs_id2str(req->rq_peer), req->rq_xid);
1467 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1468 lustre_msg_get_opc(req->rq_reqmsg) == obd_fail_val) {
1469 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1470 obd_fail_val, req->rq_xid);
1475 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1476 CERROR("wrong packet type received (type=%u) from %s\n",
1477 lustre_msg_get_type(req->rq_reqmsg),
1478 libcfs_id2str(req->rq_peer));
1482 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1485 req->rq_bulk_write = 1;
1489 req->rq_bulk_read = 1;
1493 CDEBUG(D_NET, "got req "LPU64"\n", req->rq_xid);
1495 req->rq_export = class_conn2export(
1496 lustre_msg_get_handle(req->rq_reqmsg));
1497 if (req->rq_export) {
1498 rc = ptlrpc_check_req(req);
1500 rc = sptlrpc_target_export_check(req->rq_export, req);
1502 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1503 "illegal security flavor,");
1508 ptlrpc_update_export_timer(req->rq_export, 0);
1511 /* req_in handling should/must be fast */
1512 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1513 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1514 cfs_time_sub(cfs_time_current_sec(),
1515 req->rq_arrival_time.tv_sec));
1517 /* Set rpc server deadline and add it to the timed list */
1518 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1519 MSGHDR_AT_SUPPORT) ?
1520 /* The max time the client expects us to take */
1521 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1522 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1523 if (unlikely(deadline == 0)) {
1524 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1528 ptlrpc_at_add_timed(req);
1529 rc = ptlrpc_hpreq_init(svc, req);
1533 /* Move it over to the request processing queue */
1534 rc = ptlrpc_server_request_add(svc, req);
1537 cfs_waitq_signal(&svc->srv_waitq);
1541 cfs_spin_lock(&svc->srv_lock);
1542 svc->srv_n_queued_reqs--;
1543 svc->srv_n_active_reqs++;
1544 cfs_spin_unlock(&svc->srv_lock);
1545 ptlrpc_server_finish_request(req);
1551 * Main incoming request handling logic.
1552 * Calls handler function from service to do actual processing.
1555 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1556 struct ptlrpc_thread *thread)
1558 struct obd_export *export = NULL;
1559 struct ptlrpc_request *request;
1560 struct timeval work_start;
1561 struct timeval work_end;
1569 cfs_spin_lock(&svc->srv_lock);
1571 /* !@%$# liblustre only has 1 thread */
1572 if (cfs_atomic_read(&svc->srv_n_difficult_replies) != 0) {
1573 cfs_spin_unlock(&svc->srv_lock);
1577 request = ptlrpc_server_request_get(svc, 0);
1578 if (request == NULL) {
1579 cfs_spin_unlock(&svc->srv_lock);
1583 opc = lustre_msg_get_opc(request->rq_reqmsg);
1584 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1585 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1586 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1587 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1589 if (unlikely(fail_opc)) {
1590 if (request->rq_export && request->rq_ops) {
1591 cfs_spin_unlock(&svc->srv_lock);
1592 OBD_FAIL_TIMEOUT(fail_opc, 4);
1593 cfs_spin_lock(&svc->srv_lock);
1594 request = ptlrpc_server_request_get(svc, 0);
1595 if (request == NULL) {
1596 cfs_spin_unlock(&svc->srv_lock);
1602 cfs_list_del_init(&request->rq_list);
1603 svc->srv_n_queued_reqs--;
1604 svc->srv_n_active_reqs++;
1608 /* The phase is changed under the lock here because we need to know
1609 * the request is under processing (see ptlrpc_hpreq_reorder()). */
1610 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1611 cfs_spin_unlock(&svc->srv_lock);
1613 ptlrpc_hpreq_fini(request);
1615 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1616 libcfs_debug_dumplog();
1618 cfs_gettimeofday(&work_start);
1619 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1620 if (likely(svc->srv_stats != NULL)) {
1621 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1623 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1624 svc->srv_n_queued_reqs);
1625 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1626 svc->srv_n_active_reqs);
1627 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1628 at_get(&svc->srv_at_estimate));
1631 rc = lu_context_init(&request->rq_session,
1632 LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
1634 CERROR("Failure to initialize session: %d\n", rc);
1637 request->rq_session.lc_thread = thread;
1638 request->rq_session.lc_cookie = 0x5;
1639 lu_context_enter(&request->rq_session);
1641 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1643 request->rq_svc_thread = thread;
1645 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1647 if (likely(request->rq_export)) {
1648 if (unlikely(ptlrpc_check_req(request)))
1650 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1651 export = class_export_rpc_get(request->rq_export);
1654 /* Discard requests queued for longer than the deadline.
1655 The deadline is increased if we send an early reply. */
1656 if (cfs_time_current_sec() > request->rq_deadline) {
1657 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1658 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1659 libcfs_id2str(request->rq_peer),
1660 cfs_time_sub(request->rq_deadline,
1661 request->rq_arrival_time.tv_sec),
1662 cfs_time_sub(cfs_time_current_sec(),
1663 request->rq_deadline));
1664 goto put_rpc_export;
1667 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1668 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1669 (request->rq_export ?
1670 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1671 (request->rq_export ?
1672 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1673 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1674 libcfs_id2str(request->rq_peer),
1675 lustre_msg_get_opc(request->rq_reqmsg));
1677 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1678 OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1680 rc = svc->srv_handler(request);
1682 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1686 class_export_rpc_put(export);
1688 lu_context_exit(&request->rq_session);
1689 lu_context_fini(&request->rq_session);
1691 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1692 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1693 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1694 " client may timeout.",
1695 request->rq_xid, cfs_time_sub(request->rq_deadline,
1696 request->rq_arrival_time.tv_sec),
1697 cfs_time_sub(cfs_time_current_sec(),
1698 request->rq_deadline));
1701 cfs_gettimeofday(&work_end);
1702 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1703 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1704 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1705 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1707 (request->rq_export ?
1708 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1709 (request->rq_export ?
1710 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1711 lustre_msg_get_status(request->rq_reqmsg),
1713 libcfs_id2str(request->rq_peer),
1714 lustre_msg_get_opc(request->rq_reqmsg),
1716 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1717 (request->rq_repmsg ?
1718 lustre_msg_get_transno(request->rq_repmsg) :
1719 request->rq_transno),
1721 (request->rq_repmsg ?
1722 lustre_msg_get_status(request->rq_repmsg) : -999));
1723 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1724 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1725 int opc = opcode_offset(op);
1726 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1727 LASSERT(opc < LUSTRE_MAX_OPCODES);
1728 lprocfs_counter_add(svc->srv_stats,
1729 opc + EXTRA_MAX_OPCODES,
1733 if (unlikely(request->rq_early_count)) {
1734 DEBUG_REQ(D_ADAPTTO, request,
1735 "sent %d early replies before finishing in "
1737 request->rq_early_count,
1738 cfs_time_sub(work_end.tv_sec,
1739 request->rq_arrival_time.tv_sec));
1743 cfs_spin_lock(&svc->srv_lock);
1746 cfs_spin_unlock(&svc->srv_lock);
1747 ptlrpc_server_finish_request(request);
1753 * An internal function to process a single reply state object.
1756 ptlrpc_handle_rs (struct ptlrpc_reply_state *rs)
1758 struct ptlrpc_service *svc = rs->rs_service;
1759 struct obd_export *exp;
1760 struct obd_device *obd;
1765 exp = rs->rs_export;
1768 LASSERT (rs->rs_difficult);
1769 LASSERT (rs->rs_scheduled);
1770 LASSERT (cfs_list_empty(&rs->rs_list));
1772 cfs_spin_lock (&exp->exp_lock);
1773 /* Noop if removed already */
1774 cfs_list_del_init (&rs->rs_exp_list);
1775 cfs_spin_unlock (&exp->exp_lock);
1777 /* The disk commit callback holds exp_uncommitted_replies_lock while it
1778 * iterates over newly committed replies, removing them from
1779 * exp_uncommitted_replies. It then drops this lock and schedules the
1780 * replies it found for handling here.
1782 * We can avoid contention for exp_uncommitted_replies_lock between the
1783 * HRT threads and further commit callbacks by checking rs_committed
1784 * which is set in the commit callback while it holds both
1785 * rs_lock and exp_uncommitted_reples.
1787 * If we see rs_committed clear, the commit callback _may_ not have
1788 * handled this reply yet and we race with it to grab
1789 * exp_uncommitted_replies_lock before removing the reply from
1790 * exp_uncommitted_replies. Note that if we lose the race and the
1791 * reply has already been removed, list_del_init() is a noop.
1793 * If we see rs_committed set, we know the commit callback is handling,
1794 * or has handled this reply since store reordering might allow us to
1795 * see rs_committed set out of sequence. But since this is done
1796 * holding rs_lock, we can be sure it has all completed once we hold
1797 * rs_lock, which we do right next.
1799 if (!rs->rs_committed) {
1800 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
1801 cfs_list_del_init(&rs->rs_obd_list);
1802 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
1805 cfs_spin_lock(&rs->rs_lock);
1807 been_handled = rs->rs_handled;
1810 nlocks = rs->rs_nlocks; /* atomic "steal", but */
1811 rs->rs_nlocks = 0; /* locks still on rs_locks! */
1813 if (nlocks == 0 && !been_handled) {
1814 /* If we see this, we should already have seen the warning
1815 * in mds_steal_ack_locks() */
1816 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1819 rs->rs_xid, rs->rs_transno, rs->rs_opc,
1820 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1823 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1824 cfs_spin_unlock(&rs->rs_lock);
1826 if (!been_handled && rs->rs_on_net) {
1827 LNetMDUnlink(rs->rs_md_h);
1828 /* Ignore return code; we're racing with
1832 while (nlocks-- > 0)
1833 ldlm_lock_decref(&rs->rs_locks[nlocks],
1834 rs->rs_modes[nlocks]);
1836 cfs_spin_lock(&rs->rs_lock);
1839 rs->rs_scheduled = 0;
1841 if (!rs->rs_on_net) {
1843 cfs_spin_unlock(&rs->rs_lock);
1845 class_export_put (exp);
1846 rs->rs_export = NULL;
1847 ptlrpc_rs_decref (rs);
1848 cfs_atomic_dec (&svc->srv_outstanding_replies);
1849 if (cfs_atomic_dec_and_test(&svc->srv_n_difficult_replies) &&
1850 svc->srv_is_stopping)
1851 cfs_waitq_broadcast(&svc->srv_waitq);
1855 /* still on the net; callback will schedule */
1856 cfs_spin_unlock(&rs->rs_lock);
1863 * Check whether given service has a reply available for processing
1866 * \param svc a ptlrpc service
1867 * \retval 0 no replies processed
1868 * \retval 1 one reply processed
1871 ptlrpc_server_handle_reply(struct ptlrpc_service *svc)
1873 struct ptlrpc_reply_state *rs = NULL;
1876 cfs_spin_lock(&svc->srv_lock);
1877 if (!cfs_list_empty(&svc->srv_reply_queue)) {
1878 rs = cfs_list_entry(svc->srv_reply_queue.prev,
1879 struct ptlrpc_reply_state,
1881 cfs_list_del_init(&rs->rs_list);
1883 cfs_spin_unlock(&svc->srv_lock);
1885 ptlrpc_handle_rs(rs);
1889 /* FIXME make use of timeout later */
1891 liblustre_check_services (void *arg)
1893 int did_something = 0;
1895 cfs_list_t *tmp, *nxt;
1898 /* I'm relying on being single threaded, not to have to lock
1899 * ptlrpc_all_services etc */
1900 cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1901 struct ptlrpc_service *svc =
1902 cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
1904 if (svc->srv_threads_running != 0) /* I've recursed */
1907 /* service threads can block for bulk, so this limits us
1908 * (arbitrarily) to recursing 1 stack frame per service.
1909 * Note that the problem with recursion is that we have to
1910 * unwind completely before our caller can resume. */
1912 svc->srv_threads_running++;
1915 rc = ptlrpc_server_handle_req_in(svc);
1916 rc |= ptlrpc_server_handle_reply(svc);
1917 rc |= ptlrpc_at_check_timed(svc);
1918 rc |= ptlrpc_server_handle_request(svc, NULL);
1919 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1920 did_something |= rc;
1923 svc->srv_threads_running--;
1926 RETURN(did_something);
1928 #define ptlrpc_stop_all_threads(s) do {} while (0)
1930 #else /* __KERNEL__ */
1933 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1935 int avail = svc->srv_nrqbd_receiving;
1936 int low_water = test_req_buffer_pressure ? 0 :
1937 svc->srv_nbuf_per_group/2;
1939 /* NB I'm not locking; just looking. */
1941 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1942 * allowed the request history to grow out of control. We could put a
1943 * sanity check on that here and cull some history if we need the
1946 if (avail <= low_water)
1947 ptlrpc_grow_req_bufs(svc);
1950 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1955 ptlrpc_retry_rqbds(void *arg)
1957 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1959 svc->srv_rqbd_timeout = 0;
1960 return (-ETIMEDOUT);
1964 * Status bits to pass todo info from
1965 * ptlrpc_main_check_event to ptlrpc_main.
1967 #define PTLRPC_MAIN_STOPPING 0x01
1968 #define PTLRPC_MAIN_IN_REQ 0x02
1969 #define PTLRPC_MAIN_ACTIVE_REQ 0x04
1970 #define PTLRPC_MAIN_CHECK_TIMED 0x08
1971 #define PTLRPC_MAIN_REPOST 0x10
1974 * A container to share per-thread status variables between
1975 * ptlrpc_main_check_event and ptlrpc_main functions.
1977 struct ptlrpc_main_check_s {
1978 /** todo info for the ptrlrpc_main */
1980 /** is this thread counted as running or not? */
1985 * Check whether current service thread has work to do.
1987 static int ptlrpc_main_check_event(struct ptlrpc_thread *t,
1988 struct ptlrpc_main_check_s *status)
1990 struct ptlrpc_service *svc = t->t_svc;
1995 /* check the stop flags w/o any locking to make all
1996 * concurrently running threads stop faster. */
1997 if (unlikely((t->t_flags & SVC_STOPPING) ||
1998 svc->srv_is_stopping)) {
1999 status->todo |= PTLRPC_MAIN_STOPPING;
2003 cfs_spin_lock(&svc->srv_lock);
2004 /* ptlrpc_server_request_pending() needs this thread to be
2005 * counted as running. */
2006 if (!status->running) {
2007 svc->srv_threads_running++;
2008 status->running = 1;
2010 /* Process all incoming reqs before handling any */
2011 if (!cfs_list_empty(&svc->srv_req_in_queue)) {
2012 status->todo |= PTLRPC_MAIN_IN_REQ;
2014 /* Don't handle regular requests in the last thread, in order
2015 * to handle any incoming reqs, early replies, etc. */
2016 if (ptlrpc_server_request_pending(svc, 0) &&
2017 svc->srv_threads_running <= svc->srv_threads_started - 1) {
2018 status->todo |= PTLRPC_MAIN_ACTIVE_REQ;
2020 if (svc->srv_at_check) {
2021 status->todo |= PTLRPC_MAIN_CHECK_TIMED;
2023 if (!cfs_list_empty(&svc->srv_idle_rqbds) &&
2024 svc->srv_rqbd_timeout == 0) {
2025 status->todo |= PTLRPC_MAIN_REPOST;
2027 /* count this thread as not running if it is going to sleep in
2028 * the outer wait event */
2029 if (!status->todo) {
2030 svc->srv_threads_running--;
2031 status->running = 0;
2033 cfs_spin_unlock(&svc->srv_lock);
2035 RETURN(status->todo);
2039 * Main thread body for service threads.
2040 * Waits in a loop waiting for new requests to process to appear.
2041 * Every time an incoming requests is added to its queue, a waitq
2042 * is woken up and one of the threads will handle it.
2044 static int ptlrpc_main(void *arg)
2046 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
2047 struct ptlrpc_service *svc = data->svc;
2048 struct ptlrpc_thread *thread = data->thread;
2049 struct obd_device *dev = data->dev;
2050 struct ptlrpc_reply_state *rs;
2051 struct ptlrpc_main_check_s st;
2052 #ifdef WITH_GROUP_INFO
2053 cfs_group_info_t *ginfo = NULL;
2056 int counter = 0, rc = 0;
2059 thread->t_pid = cfs_curproc_pid();
2060 cfs_daemonize_ctxt(data->name);
2062 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
2063 /* we need to do this before any per-thread allocation is done so that
2064 * we get the per-thread allocations on local node. bug 7342 */
2065 if (svc->srv_cpu_affinity) {
2068 for (cpu = 0, num_cpu = 0; cpu < cfs_num_possible_cpus();
2070 if (!cfs_cpu_online(cpu))
2072 if (num_cpu == thread->t_id % cfs_num_online_cpus())
2076 cfs_set_cpus_allowed(cfs_current(),
2077 node_to_cpumask(cpu_to_node(cpu)));
2081 #ifdef WITH_GROUP_INFO
2082 ginfo = cfs_groups_alloc(0);
2088 cfs_set_current_groups(ginfo);
2089 cfs_put_group_info(ginfo);
2092 if (svc->srv_init != NULL) {
2093 rc = svc->srv_init(thread);
2098 rc = lu_context_init(&env.le_ctx,
2099 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2103 thread->t_env = &env;
2104 env.le_ctx.lc_thread = thread;
2105 env.le_ctx.lc_cookie = 0x6;
2107 /* Alloc reply state structure for this one */
2108 OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
2114 cfs_spin_lock(&svc->srv_lock);
2115 /* SVC_STOPPING may already be set here if someone else is trying
2116 * to stop the service while this new thread has been dynamically
2117 * forked. We still set SVC_RUNNING to let our creator know that
2118 * we are now running, however we will exit as soon as possible */
2119 thread->t_flags |= SVC_RUNNING;
2120 cfs_spin_unlock(&svc->srv_lock);
2123 * wake up our creator. Note: @data is invalid after this point,
2124 * because it's allocated on ptlrpc_start_thread() stack.
2126 cfs_waitq_signal(&thread->t_ctl_waitq);
2128 thread->t_watchdog = lc_watchdog_add(CFS_GET_TIMEOUT(svc), NULL, NULL);
2130 cfs_spin_lock(&svc->srv_lock);
2131 cfs_list_add(&rs->rs_list, &svc->srv_free_rs_list);
2132 cfs_spin_unlock(&svc->srv_lock);
2133 cfs_waitq_signal(&svc->srv_free_rs_waitq);
2135 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2136 svc->srv_threads_running);
2138 /* XXX maintain a list of all managed devices: insert here */
2143 while (!(st.todo & PTLRPC_MAIN_STOPPING)) {
2144 /* Don't exit while there are replies to be handled */
2145 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
2146 ptlrpc_retry_rqbds, svc);
2148 lc_watchdog_disable(thread->t_watchdog);
2152 l_wait_event_exclusive (svc->srv_waitq,
2153 ptlrpc_main_check_event(thread, &st),
2156 lc_watchdog_touch(thread->t_watchdog, CFS_GET_TIMEOUT(svc));
2158 ptlrpc_check_rqbd_pool(svc);
2160 if (svc->srv_threads_started < svc->srv_threads_max &&
2161 svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))
2162 /* Ignore return code - we tried... */
2163 ptlrpc_start_thread(dev, svc);
2165 /* Process all incoming reqs before handling any */
2166 if (st.todo & PTLRPC_MAIN_IN_REQ) {
2167 ptlrpc_server_handle_req_in(svc);
2168 /* but limit ourselves in case of flood */
2169 if (counter++ < 1000)
2173 if (st.todo & PTLRPC_MAIN_CHECK_TIMED) {
2174 ptlrpc_at_check_timed(svc);
2176 if (st.todo & PTLRPC_MAIN_ACTIVE_REQ) {
2177 lu_context_enter(&env.le_ctx);
2178 ptlrpc_server_handle_request(svc, thread);
2179 lu_context_exit(&env.le_ctx);
2181 if ((st.todo & PTLRPC_MAIN_REPOST) &&
2182 ptlrpc_server_post_idle_rqbds(svc) < 0) {
2183 /* I just failed to repost request buffers.
2184 * Wait for a timeout (unless something else
2185 * happens) before I try again */
2186 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
2187 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
2188 svc->srv_nrqbd_receiving);
2192 lc_watchdog_delete(thread->t_watchdog);
2193 thread->t_watchdog = NULL;
2197 * deconstruct service specific state created by ptlrpc_start_thread()
2199 if (svc->srv_done != NULL)
2200 svc->srv_done(thread);
2202 lu_context_fini(&env.le_ctx);
2204 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2205 thread, thread->t_pid, thread->t_id, rc);
2207 cfs_spin_lock(&svc->srv_lock);
2209 svc->srv_threads_running--;
2211 thread->t_flags = SVC_STOPPED;
2212 cfs_waitq_signal(&thread->t_ctl_waitq);
2213 cfs_spin_unlock(&svc->srv_lock);
2218 struct ptlrpc_hr_args {
2221 struct ptlrpc_hr_service *hrs;
2224 static int hrt_dont_sleep(struct ptlrpc_hr_thread *t,
2225 cfs_list_t *replies)
2229 cfs_spin_lock(&t->hrt_lock);
2230 cfs_list_splice_init(&t->hrt_queue, replies);
2231 result = cfs_test_bit(HRT_STOPPING, &t->hrt_flags) ||
2232 !cfs_list_empty(replies);
2233 cfs_spin_unlock(&t->hrt_lock);
2238 * Main body of "handle reply" function.
2239 * It processes acked reply states
2241 static int ptlrpc_hr_main(void *arg)
2243 struct ptlrpc_hr_args * hr_args = arg;
2244 struct ptlrpc_hr_service *hr = hr_args->hrs;
2245 struct ptlrpc_hr_thread *t = &hr->hr_threads[hr_args->thread_index];
2246 char threadname[20];
2247 CFS_LIST_HEAD(replies);
2249 snprintf(threadname, sizeof(threadname),
2250 "ptlrpc_hr_%d", hr_args->thread_index);
2252 cfs_daemonize_ctxt(threadname);
2253 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
2254 cfs_set_cpus_allowed(cfs_current(),
2255 node_to_cpumask(cpu_to_node(hr_args->cpu_index)));
2257 cfs_set_bit(HRT_RUNNING, &t->hrt_flags);
2258 cfs_waitq_signal(&t->hrt_wait);
2260 while (!cfs_test_bit(HRT_STOPPING, &t->hrt_flags)) {
2262 l_cfs_wait_event(t->hrt_wait, hrt_dont_sleep(t, &replies));
2263 while (!cfs_list_empty(&replies)) {
2264 struct ptlrpc_reply_state *rs;
2266 rs = cfs_list_entry(replies.prev,
2267 struct ptlrpc_reply_state,
2269 cfs_list_del_init(&rs->rs_list);
2270 ptlrpc_handle_rs(rs);
2274 cfs_clear_bit(HRT_RUNNING, &t->hrt_flags);
2275 cfs_complete(&t->hrt_completion);
2280 static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu)
2282 struct ptlrpc_hr_thread *t = &hr->hr_threads[n];
2283 struct ptlrpc_hr_args args;
2287 args.thread_index = n;
2288 args.cpu_index = cpu;
2291 rc = cfs_kernel_thread(ptlrpc_hr_main, (void*)&args,
2292 CLONE_VM|CLONE_FILES);
2294 cfs_complete(&t->hrt_completion);
2297 l_cfs_wait_event(t->hrt_wait, cfs_test_bit(HRT_RUNNING, &t->hrt_flags));
2303 static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t)
2307 cfs_set_bit(HRT_STOPPING, &t->hrt_flags);
2308 cfs_waitq_signal(&t->hrt_wait);
2309 cfs_wait_for_completion(&t->hrt_completion);
2314 static void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs)
2319 for (n = 0; n < hrs->hr_n_threads; n++)
2320 ptlrpc_stop_hr_thread(&hrs->hr_threads[n]);
2325 static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr)
2328 int n, cpu, threads_started = 0;
2331 LASSERT(hr != NULL);
2332 LASSERT(hr->hr_n_threads > 0);
2334 for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) {
2335 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
2336 while(!cfs_cpu_online(cpu)) {
2338 if (cpu >= cfs_num_possible_cpus())
2342 rc = ptlrpc_start_hr_thread(hr, n, cpu);
2348 if (threads_started == 0) {
2349 CERROR("No reply handling threads started\n");
2352 if (threads_started < hr->hr_n_threads) {
2353 CWARN("Started only %d reply handling threads from %d\n",
2354 threads_started, hr->hr_n_threads);
2355 hr->hr_n_threads = threads_started;
2360 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
2361 struct ptlrpc_thread *thread)
2363 struct l_wait_info lwi = { 0 };
2366 CDEBUG(D_RPCTRACE, "Stopping thread [ %p : %u ]\n",
2367 thread, thread->t_pid);
2369 cfs_spin_lock(&svc->srv_lock);
2370 /* let the thread know that we would like it to stop asap */
2371 thread->t_flags |= SVC_STOPPING;
2372 cfs_spin_unlock(&svc->srv_lock);
2374 cfs_waitq_broadcast(&svc->srv_waitq);
2375 l_wait_event(thread->t_ctl_waitq,
2376 (thread->t_flags & SVC_STOPPED), &lwi);
2378 cfs_spin_lock(&svc->srv_lock);
2379 cfs_list_del(&thread->t_link);
2380 cfs_spin_unlock(&svc->srv_lock);
2382 OBD_FREE_PTR(thread);
2387 * Stops all threads of a particular service \a svc
2389 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2391 struct ptlrpc_thread *thread;
2394 cfs_spin_lock(&svc->srv_lock);
2395 while (!cfs_list_empty(&svc->srv_threads)) {
2396 thread = cfs_list_entry(svc->srv_threads.next,
2397 struct ptlrpc_thread, t_link);
2399 cfs_spin_unlock(&svc->srv_lock);
2400 ptlrpc_stop_thread(svc, thread);
2401 cfs_spin_lock(&svc->srv_lock);
2404 cfs_spin_unlock(&svc->srv_lock);
2408 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
2413 /* We require 2 threads min - see note in
2414 ptlrpc_server_handle_request */
2415 LASSERT(svc->srv_threads_min >= 2);
2416 for (i = 0; i < svc->srv_threads_min; i++) {
2417 rc = ptlrpc_start_thread(dev, svc);
2418 /* We have enough threads, don't start more. b=15759 */
2422 CERROR("cannot start %s thread #%d: rc %d\n",
2423 svc->srv_thread_name, i, rc);
2424 ptlrpc_stop_all_threads(svc);
2431 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
2433 struct l_wait_info lwi = { 0 };
2434 struct ptlrpc_svc_data d;
2435 struct ptlrpc_thread *thread;
2440 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
2441 svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
2442 svc->srv_threads_max, svc->srv_threads_running);
2444 if (unlikely(svc->srv_is_stopping))
2447 if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
2448 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2449 svc->srv_threads_started == svc->srv_threads_min - 1))
2452 OBD_ALLOC_PTR(thread);
2455 cfs_waitq_init(&thread->t_ctl_waitq);
2457 cfs_spin_lock(&svc->srv_lock);
2458 if (svc->srv_threads_started >= svc->srv_threads_max) {
2459 cfs_spin_unlock(&svc->srv_lock);
2460 OBD_FREE_PTR(thread);
2463 cfs_list_add(&thread->t_link, &svc->srv_threads);
2464 id = svc->srv_threads_started++;
2465 cfs_spin_unlock(&svc->srv_lock);
2467 thread->t_svc = svc;
2469 sprintf(name, "%s_%02d", svc->srv_thread_name, id);
2475 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
2477 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2478 * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
2480 rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
2482 CERROR("cannot start thread '%s': rc %d\n", name, rc);
2484 cfs_spin_lock(&svc->srv_lock);
2485 cfs_list_del(&thread->t_link);
2486 --svc->srv_threads_started;
2487 cfs_spin_unlock(&svc->srv_lock);
2489 OBD_FREE(thread, sizeof(*thread));
2492 l_wait_event(thread->t_ctl_waitq,
2493 thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
2495 rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
2500 int ptlrpc_hr_init(void)
2503 int n_cpus = cfs_num_online_cpus();
2504 struct ptlrpc_hr_service *hr;
2509 LASSERT(ptlrpc_hr == NULL);
2511 size = offsetof(struct ptlrpc_hr_service, hr_threads[n_cpus]);
2512 OBD_ALLOC(hr, size);
2515 for (i = 0; i < n_cpus; i++) {
2516 struct ptlrpc_hr_thread *t = &hr->hr_threads[i];
2518 cfs_spin_lock_init(&t->hrt_lock);
2519 cfs_waitq_init(&t->hrt_wait);
2520 CFS_INIT_LIST_HEAD(&t->hrt_queue);
2521 cfs_init_completion(&t->hrt_completion);
2523 hr->hr_n_threads = n_cpus;
2527 rc = ptlrpc_start_hr_threads(hr);
2529 OBD_FREE(hr, hr->hr_size);
2535 void ptlrpc_hr_fini(void)
2537 if (ptlrpc_hr != NULL) {
2538 ptlrpc_stop_hr_threads(ptlrpc_hr);
2539 OBD_FREE(ptlrpc_hr, ptlrpc_hr->hr_size);
2544 #endif /* __KERNEL__ */
2547 * Wait until all already scheduled replies are processed.
2549 static void ptlrpc_wait_replies(struct ptlrpc_service *svc)
2553 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2555 rc = l_wait_event(svc->srv_waitq, cfs_atomic_read(&svc-> \
2556 srv_n_difficult_replies) == 0,
2560 CWARN("Unexpectedly long timeout %p\n", svc);
2564 int ptlrpc_unregister_service(struct ptlrpc_service *service)
2567 struct l_wait_info lwi;
2569 struct ptlrpc_reply_state *rs, *t;
2570 struct ptlrpc_at_array *array = &service->srv_at_array;
2573 service->srv_is_stopping = 1;
2574 cfs_timer_disarm(&service->srv_at_timer);
2576 ptlrpc_stop_all_threads(service);
2577 LASSERT(cfs_list_empty(&service->srv_threads));
2579 cfs_spin_lock (&ptlrpc_all_services_lock);
2580 cfs_list_del_init (&service->srv_list);
2581 cfs_spin_unlock (&ptlrpc_all_services_lock);
2583 ptlrpc_lprocfs_unregister_service(service);
2585 /* All history will be culled when the next request buffer is
2587 service->srv_max_history_rqbds = 0;
2589 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
2591 rc = LNetClearLazyPortal(service->srv_req_portal);
2594 /* Unlink all the request buffers. This forces a 'final' event with
2595 * its 'unlink' flag set for each posted rqbd */
2596 cfs_list_for_each(tmp, &service->srv_active_rqbds) {
2597 struct ptlrpc_request_buffer_desc *rqbd =
2598 cfs_list_entry(tmp, struct ptlrpc_request_buffer_desc,
2601 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2602 LASSERT (rc == 0 || rc == -ENOENT);
2605 /* Wait for the network to release any buffers it's currently
2608 cfs_spin_lock(&service->srv_lock);
2609 rc = service->srv_nrqbd_receiving;
2610 cfs_spin_unlock(&service->srv_lock);
2615 /* Network access will complete in finite time but the HUGE
2616 * timeout lets us CWARN for visibility of sluggish NALs */
2617 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2618 cfs_time_seconds(1), NULL, NULL);
2619 rc = l_wait_event(service->srv_waitq,
2620 service->srv_nrqbd_receiving == 0,
2622 if (rc == -ETIMEDOUT)
2623 CWARN("Service %s waiting for request buffers\n",
2627 /* schedule all outstanding replies to terminate them */
2628 cfs_spin_lock(&service->srv_lock);
2629 while (!cfs_list_empty(&service->srv_active_replies)) {
2630 struct ptlrpc_reply_state *rs =
2631 cfs_list_entry(service->srv_active_replies.next,
2632 struct ptlrpc_reply_state, rs_list);
2633 cfs_spin_lock(&rs->rs_lock);
2634 ptlrpc_schedule_difficult_reply(rs);
2635 cfs_spin_unlock(&rs->rs_lock);
2637 cfs_spin_unlock(&service->srv_lock);
2639 /* purge the request queue. NB No new replies (rqbds all unlinked)
2640 * and no service threads, so I'm the only thread noodling the
2641 * request queue now */
2642 while (!cfs_list_empty(&service->srv_req_in_queue)) {
2643 struct ptlrpc_request *req =
2644 cfs_list_entry(service->srv_req_in_queue.next,
2645 struct ptlrpc_request,
2648 cfs_list_del(&req->rq_list);
2649 service->srv_n_queued_reqs--;
2650 service->srv_n_active_reqs++;
2651 ptlrpc_server_finish_request(req);
2653 while (ptlrpc_server_request_pending(service, 1)) {
2654 struct ptlrpc_request *req;
2656 req = ptlrpc_server_request_get(service, 1);
2657 cfs_list_del(&req->rq_list);
2658 service->srv_n_queued_reqs--;
2659 service->srv_n_active_reqs++;
2660 ptlrpc_hpreq_fini(req);
2661 ptlrpc_server_finish_request(req);
2663 LASSERT(service->srv_n_queued_reqs == 0);
2664 LASSERT(service->srv_n_active_reqs == 0);
2665 LASSERT(service->srv_n_history_rqbds == 0);
2666 LASSERT(cfs_list_empty(&service->srv_active_rqbds));
2668 /* Now free all the request buffers since nothing references them
2670 while (!cfs_list_empty(&service->srv_idle_rqbds)) {
2671 struct ptlrpc_request_buffer_desc *rqbd =
2672 cfs_list_entry(service->srv_idle_rqbds.next,
2673 struct ptlrpc_request_buffer_desc,
2676 ptlrpc_free_rqbd(rqbd);
2679 ptlrpc_wait_replies(service);
2681 cfs_list_for_each_entry_safe(rs, t, &service->srv_free_rs_list,
2683 cfs_list_del(&rs->rs_list);
2684 OBD_FREE(rs, service->srv_max_reply_size);
2687 /* In case somebody rearmed this in the meantime */
2688 cfs_timer_disarm(&service->srv_at_timer);
2690 if (array->paa_reqs_array != NULL) {
2691 OBD_FREE(array->paa_reqs_array,
2692 sizeof(cfs_list_t) * array->paa_size);
2693 array->paa_reqs_array = NULL;
2696 if (array->paa_reqs_count != NULL) {
2697 OBD_FREE(array->paa_reqs_count,
2698 sizeof(__u32) * array->paa_size);
2699 array->paa_reqs_count= NULL;
2702 OBD_FREE_PTR(service);
2707 * Returns 0 if the service is healthy.
2709 * Right now, it just checks to make sure that requests aren't languishing
2710 * in the queue. We'll use this health check to govern whether a node needs
2711 * to be shot, so it's intentionally non-aggressive. */
2712 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2714 struct ptlrpc_request *request;
2715 struct timeval right_now;
2721 cfs_gettimeofday(&right_now);
2723 cfs_spin_lock(&svc->srv_lock);
2724 if (!ptlrpc_server_request_pending(svc, 1)) {
2725 cfs_spin_unlock(&svc->srv_lock);
2729 /* How long has the next entry been waiting? */
2730 if (cfs_list_empty(&svc->srv_request_queue))
2731 request = cfs_list_entry(svc->srv_request_hpq.next,
2732 struct ptlrpc_request, rq_list);
2734 request = cfs_list_entry(svc->srv_request_queue.next,
2735 struct ptlrpc_request, rq_list);
2736 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2737 cfs_spin_unlock(&svc->srv_lock);
2739 if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2741 CERROR("%s: unhealthy - request has been waiting %lds\n",
2742 svc->srv_name, timediff / ONE_MILLION);